A small, stubborn truth engine

Problem

Fluent language is cheap. Feeds, models, and marketing all produce sentences that sound settled. Most tools optimize for engagement, speed, or agreeableness. Few optimize for surviving contact with evidence — and fewer still show their reasoning as a first-class object you can attack.

The rare thing worth building: something that would rather be corrected than look confident.

Design principles

1. Correct over agreeable. If a claim fails, say so without soft padding that hides the failure.
2. Show the work. Arguments are graphs, not monologues: claim → support → gaps → status.
3. Confidence is earned and revocable. Thin evidence means a short confidence bar — visually, not just in fine print.
4. Public correction. When the engine was wrong, the record updates in the open. Silent rewrites are a bug.
5. Try to kill the idea. A first-class mode that attacks the claim before anyone falls in love with it.
6. Unknown is a valid output. “We don’t know yet” beats a fake synthesis.

Status labels

Every claim node should wear one or more of these. Labels can stack.

• Solid — primary sources checked; claim survives stated tests; confidence still limited by what was measured.
• Rumor — circulating assertion with weak or circular provenance.
• Marketing — persuasion language or incentive shape dominates the packaging.
• Unknown — not enough evidence yet; the honest default.
• Broken — failed a check, contradicted by stronger evidence, or internally incoherent. “This broke when we checked.”

The sketch on this page will not award solid. That label requires real primary sources — not pattern matching in a browser.

Intended pipeline

1. Ingest — accept a claim from news, papers, posts, or notes. Normalize it into a falsifiable statement when possible.
2. Decompose — split into sub-claims, quantities, causal links, and hidden assumptions.
3. Stress-test — seek primary sources, contradictions, incentives, and measurability. Prefer originals over summaries of summaries.
4. Graph — render support, attacks, and gaps as a map people can inspect without trusting a black-box paragraph.
5. Score with humility — confidence shrinks when evidence is thin, contested, or misaligned with the claim’s scope.
6. Publish the trail — what was checked, what failed, what remains open. Corrections append; they don’t vanish.

What we refuse to build

• A vibes chatbot that answers with unearned certainty.
• A “trust us” score with no inspectable chain of work.
• Silent edits that make past mistakes disappear.
• Engagement features that reward heat over testability.

Interactive sketch — structure only

Paste a claim. The sketch looks for absolute language, causal leaps, numbers without provenance, marketing tone, and missing hedges. It returns assumptions, tests, and a “try to kill it” list.

Runs entirely in your browser. No accounts. No claim storage by this page. No web lookup. If it ever pretends to have checked a paper, that would be a regression — report it as a bug in spirit.

Try an example:

Roadmap (public, provisional)

1. Now — principles, labels, and this structural sketch.
2. Next — richer claim graphs (multiple nodes, explicit support/attack edges) still without fake verification.
3. Later — optional source attachment workflow: humans or tools add citations; the engine scores provenance quality and still allows “unknown.”
4. Much later — public correction log for claims that changed status over time.

Simulation Universe waits its turn. Depth on truth first; joy second.

Open questions

• How much structure can stay useful before the UI becomes a chore?
• What’s the minimum honest interface for “we checked X and it broke”?
• When sources conflict, how do we show disagreement without false balance or false resolution?